The Eocene–Oligocene transition (EOT) was a climate shift from a largely ice-free greenhouse world to an icehouse climate, involving the first major glaciation of Antarctica and global cooling occurring ∼ 34 million years ago (Ma) and lasting ∼ 790 kyr. The change is marked by a global shift in deep-sea δ18O representing a combination of deep-ocean cooling and growth in land ice volume. At the same time, multiple independent proxies for ocean tempera- ture indicate sea surface cooling, and major changes in global fauna and flora record a shift toward more cold-climate- adapted species. The two principal suggested explanations of this transition are a decline in atmospheric CO2 and changes to ocean gateways, while orbital forcing likely influenced the precise timing of the glaciation. Here we review and synthesise proxy evidence of palaeogeography, temperature, ice sheets, ocean circulation and CO2 change from the marine and terrestrial realms. Furthermore, we quantitatively com- pare proxy records of change to an ensemble of climate model simulations of temperature change across the EOT. The simulations compare three forcing mechanisms across the EOT: CO2 decrease, palaeogeographic changes and ice sheet growth. Our model ensemble results demonstrate the need for a global cooling mechanism beyond the imposition of an ice sheet or palaeogeographic changes. We find that CO2 forcing involving a large decrease in CO2 of ca. 40 % (∼ 325 ppm drop) provides the best fit to the available proxy evidence, with ice sheet and palaeogeographic changes play- ing a secondary role. While this large decrease is consistent with some CO2 proxy records (the extreme endmember of decrease), the positive feedback mechanisms on ice growth are so strong that a modest CO2 decrease beyond a critical threshold for ice sheet initiation is well capable of triggering rapid ice sheet growth. Thus, the amplitude of CO2 decrease signalled by our data–model comparison should be consid- ered an upper estimate and perhaps artificially large, not least because the current generation of climate models do not in- clude dynamic ice sheets and in some cases may be under- sensitive to CO2 forcing. The model ensemble also cannot exclude the possibility that palaeogeographic changes could have triggered a reduction in CO2.